a. Lack of data and presence of bias

While efforts to implement lab automation using data-driven approaches and digital laboratories are actively progressing, challenges such as a lack of high-quality datasets and the predominance of data biased toward specific materials and experimental conditions remain critical challenges. In such cases, it becomes inevitable to rely on existing data for analysis instead of synthesizing new materials, thereby exacerbating the bias. These problems adversely affect AI model training and optimization, ultimately reducing prediction accuracy.

b. Treatment of diverse materials

In the design and synthesis of nanomaterials, ceramics, and metallic materials, careful consideration of the complex interactions between material composition and external synthesis conditions is essential. The properties of these materials often depend significantly on subtle changes in external factors such as temperature, pressure, and chemical environment, in addition to their microscopic composition and structure. Therefore, accurately understanding and modeling the interactions between controllable variables (e.g., synthesis temperature, reaction time, catalyst quantity) and resulting material properties (e.g., strength, thermal conductivity, electrical characteristics) remains an unresolved challenge. Addressing this challenge is critically important in materials science, as it directly impacts the efficient design of novel materials and the development of high-performance materials.

c. Limitations of computational resources

Molecular dynamics simulations and first-principles calculations are widely utilized as powerful methods for predicting the properties of novel materials and assisting in their design. However, these methods often require substantial computational resources, with high computational costs and long processing times posing barriers to practical implementation. Additionally, while these simulation techniques are generally well-suited for making predictions based on physical models, they are less effective for inverse analytical approaches aimed at optimizing target variables (e.g., strength, conductivity, thermal conductivity). Consequently, it remains challenging to automatically explore the optimal compositions, structures, and external conditions necessary to achieve desired material properties, often necessitating manual trial-and-error. To overcome these limitations, integrating AI and data-driven approaches has become essential.

d. Interpretation of results

Explaining the rationale behind prediction outcomes is vital for ensuring scientific reliability. In the field of materials science, the interpretation of prediction outcomes plays a particularly critical role. This is because accurately understanding the results and uncovering the underlying causal relationships and mechanisms can provide new insights for material design and process optimization. Conversely, when predictions function as a “black box” and the reasoning behind them cannot be clearly explained, it not only undermines the reliability of applying the results but also poses significant challenges to improving the predictive model and expanding its applications. For these reasons, the introduction of methods to enhance explainability is considered crucial. Explainable predictive models not only greatly enhance their practicality as complements to experimental results and as guides for new material discovery but also serve as key enablers of acceptance in both academic and industrial applications.